Hierarchical Manycore Resource Management Framework using Control Processors

Abstract

As the number of cores integrated into a single manycore processor chip increases, the amount of computation for runtime scheduling becomes prohibitively large. In this paper, we present an efficient manycore resource management scheme with consideration of both runtime scheduling overhead and power consumption exploiting architecture support of tiny control processors. By using hierarchically placed control processors on a manycore processor, we distribute dynamic scheduling and perform efficient hierarchical load-balancing. Moreover, to utilize the core resources with energy efficiency, the proposed framework dynamically adjusts the core resource allocation for simultaneously running applications based on applications performance characteristics. Without the use of processing cores and memory resources, we perform dynamic work scheduling and space-shared scheduling for multiple parallel applications. The evaluation of our approach shows that the proposed scheduling framework achieves signifi- cant system performance improvement of 66% compared to scheduling using the processing cores. In addition, it presents efficient space-shared scheduling for parallel applications in multiprogrammed environments under a given power-budget.